Improving Query Quality

1. Query Rewriting (Query Reformulation)

Idea: Rewrite the user’s vague or incomplete query into a complete, rich query.

How:

• Use an LLM to rephrase
• Use a simple prompt: “Rewrite the query to maximize retrieval relevance”

Example:

User Query	Rewritten Query
“ai healthcare”	“What are the different applications of Artificial Intelligence in the healthcare sector, such as diagnosis, treatment, and patient monitoring?”

---

2. Context Injection (Conversation Memory)

Idea: Inject previous conversation history into the current query to add missing context.

How:

• Use ConversationBufferMemory (LangChain)
• Add last n turns of chat

Example:

Context	User Query	Final Query
Earlier discussed “AI in education”	“And healthcare?”	“What are the applications of AI in healthcare, similar to its uses in education?”

---

3. Query Expansion (Adding Synonyms/Related Terms)

Idea: Expand the query by adding synonyms or related concepts to catch more matches during retrieval.

How:

• Use ontology (thesaurus/wordnet)
• Embed expansion manually

Example:

User Query	Expanded Query
“cancer diagnosis AI”	“cancer diagnosis OR oncology diagnosis OR tumor detection using Artificial Intelligence OR Machine Learning”

---

4. Multi-Query Generation (Multiple Query Variants)

Idea: Generate multiple rephrased queries and perform retrieval for all of them to maximize coverage.

How:

• Use LLM to generate 3–5 variations
• Retrieve for each variant and merge results

Example:

User Query	Query Variants
“uses of AI in hospitals”	“applications of AI in medical field”, “how AI helps hospitals”, “AI in healthcare diagnostics”

(LangChain has MultiQueryRetriever ready for this.)

---

5. Semantic Search instead of Keyword Search

Idea: Search by meaning, not words. Even if the query is vague, embedding similarity retrieves relevant content.

How:

• Use embedding models: OpenAI, HuggingFace, Vertex AI, etc.
• Store documents in vector DB

Example:

User Query	Retrieved
“machine thinking”	Document about “Artificial Intelligence” (even if “machine thinking” isn’t explicitly written)

---

6. Fallback to Generative Answers (if retrieval fails)

Idea: If retrieval gives bad results (empty, irrelevant), fall back to pure LLM generation based on user query.

How:

• Detect low retrieval scores
• Trigger direct LLM generation with prompt “Based on your knowledge, answer…”

Example:

User Query	Fallback
“How does AI taste food?” (irrelevant or no doc)	LLM says: “Currently, AI systems can simulate aspects of taste using chemical sensors but cannot physically taste like humans.”

---

7. Query Classification and Routing

Idea: Classify user intent (e.g., FAQ, how-to, troubleshooting) and route to specialized retrieval logic.

How:

• Build intent classifiers (tiny LLM or fine-tuned classifier)
• Have different retrievers or vector indexes

Example:

User Query	Detected Intent	Retrieval
“how do I reset password?”	Troubleshooting	Retrieve from “Help Articles” index
“what is AI?”	Definition	Retrieve from “Knowledge Base”

---

Summary Table of Techniques

Technique	Purpose	Tools/Methods
Query Rewriting	Make query more complete	LLMs (Prompt Engineering)
Context Injection	Bring chat history into current query	LangChain Memories
Query Expansion	Add synonyms/related terms	Ontologies, LLM
Multi-Query Generation	Create multiple versions	LangChain MultiQueryRetriever
Semantic Search	Retrieve by meaning, not exact words	Vector Databases
Fallbacks	Ensure answers even if retrieval fails	Retrieval Confidence + Direct LLM
Query Classification	Route to specific pipelines	Classifiers (zero-shot or fine-tuned models)

---

Streamlit

• Refer Here for streamlit and Refer Here for setup and installation
• Create a new folder and activate virtual environment
• install streamlit pip install streamlit and create a requirements.txt pip freeze > requirements.txt
• Streamlit helps in building simple UI without html or css or javascript
• Create a file app.py with following content

import streamlit as st

st.title("RAG UI Prototype")

• Now run the application from terminal using streamlit run app.py
• Refer Here for widgets and Refer Here for showing progress.
• Now to debug streamlit lets add the following launch configuration in vscode

{
    "name": "Streamlit Debug",
    "type": "debugpy",
    "request": "launch",
    "module": "streamlit",
    "args": [
        "run",
        "${file}"
    ]
}

• Complete launch configuration

{
    // Use IntelliSense to learn about possible attributes.
    // Hover to view descriptions of existing attributes.
    // For more information, visit: https://go.microsoft.com/fwlink/?linkid=830387
    "version": "0.2.0",
    "configurations": [

        {
            "name": "Streamlit Debug",
            "type": "debugpy",
            "request": "launch",
            "module": "streamlit",
            "args": [
                "run",
                "${file}"
            ]
        }
    ]
}

• Now for rest of the widgets or other code look into the classroom video and github code links shared
• Refer Here for streamlit app linked to textbook rag.